Toward Efficient Privacy-Preserving Classification and Retrieval of Encrypted Images Using Deep Convolutional Neural Networks
Authors:
, ,
Journal: Multidisciplinary International Research Journal
Abstract
Retrieving and classifying encrypted data, particularly images, poses significant challenges
for privacy-preserving computing. Images play a critical role in numerous applications,
including medical imaging, military intelligence, and surveillance systems, where robust
privacy measures are essential to protect confidential information from unauthorised
access. In this paper, we present a dual-purpose framework for both classification and
content-based retrieval of encrypted images using a deep convolutional neural network
(DCNN). Our encryption scheme combines the chaotic Arnold matrix transformation with
XOR encryption using cryptographically generated 256-bit keys to provide a lightweight
privacy-obfuscation mechanism. The DCNN model, trained directly on encrypted image
representations, learns discriminative features from the encrypted domain without requiring
decryption. For the retrieval task, we extract 512-dimensional feature vectors from the
penultimate fully connected layer and employ cosine similarity to rank database images
against a query. The proposed framework was evaluated on two benchmark datasets:
MedNIST (58,954 medical images across six categories) and MNIST (70,000 handwritten
digits across ten classes). Classification results achieved 99.17% weighted-average F1-
score on MedNIST and 94.08% weighted-average F1-score on MNIST encrypted images.
Retrieval performance yielded a mean average precision (mAP) of 0.9834 on MedNIST
and 0.9127 on MNIST. Comprehensive ablation studies across four encryption conditions
and four model architectures, alongside security analysis including entropy, correlation,
NPCR, and UACI metrics, validate both the effectiveness and the privacy characteristics of
the proposed approach.
Keywords: deep learning, convolutional neural network (CNN), encrypted image, content-based image retrieval, chaotic encryption, Arnold transform, privacy-preserving computation, image classification, medical image analysis